通过 Keras 2.4.3 和 Tensorflow 2.2 从 SavedModel 中提取特征

Question

I would like to extract the features from the final dense layer in my CNN model. However, I am very conflicted with all of the google research I have done. There are so many different methods with Tensorflow and I am struggling to get something to work.

I have sucessfully trained a model on CIFAR10. I have saved the model to a directory and have a saved_model.pb file. I have visualized the model via tensorboard but not entirely sure of the name of my final layer. The visualization seems a bit confusing.

How can I proceed to extract these features? I want to use them for a t-SNE analysis.

I am trying to load the pb graph with gfile but not sure if this is the correct approach.Thank you.

import tensorflow.compat.v1 as tf
tf.disable_v2_behavior()
from tensorflow.python.platform import gfile


pb_graph_file = '../data/processed/saved_models/saved_model.pb'

f = gfile.GFile(pb_graph_file, 'rb')
graph_def = tf.GraphDef()
f.close()

My Keras Sequential model looks like:

    """
    This is the CNN model's architecture
    """
    weight_decay = 1e-4
    model = Sequential()
    model.add(Conv2D(32, (3, 3), activation = 'relu', kernel_initializer = 'he_normal', kernel_regularizer = l2(weight_decay), padding = 'same', input_shape = (32, 32, 3)))
    model.add(BatchNormalization())
    model.add(Conv2D(32, (3, 3), activation = 'relu', kernel_initializer = 'he_normal', kernel_regularizer = l2(weight_decay), padding = 'same'))
    model.add(BatchNormalization())
    model.add(MaxPooling2D((2, 2)))
    model.add(Dropout(0.2))

    model.add(Conv2D(64, (3, 3), activation = 'relu', kernel_initializer = 'he_normal', kernel_regularizer = l2(weight_decay), padding='same'))
    model.add(BatchNormalization())
    model.add(Conv2D(64, (3, 3), activation = 'relu', kernel_initializer = 'he_normal', kernel_regularizer = l2(weight_decay), padding='same'))
    model.add(BatchNormalization())
    model.add(MaxPooling2D((2, 2)))
    model.add(Dropout(0.3))

    model.add(Conv2D(128, (3, 3), activation = 'relu', kernel_initializer = 'he_normal', kernel_regularizer = l2(weight_decay), padding='same'))
    model.add(BatchNormalization())
    model.add(Conv2D(128, (3, 3), activation = 'relu', kernel_initializer = 'he_normal', kernel_regularizer = l2(weight_decay), padding='same'))
    model.add(BatchNormalization())
    model.add(MaxPooling2D((2, 2)))
    model.add(Dropout(0.4))

    # model.add(Conv2D(256, (3, 3), activation = 'relu', kernel_initializer = 'he_uniform', kernel_regularizer = l2(weight_decay), padding='same'))
    # model.add(Conv2D(256, (3, 3), activation = 'relu', kernel_initializer = 'he_uniform', kernel_regularizer = l2(weight_decay), padding='same'))
    # model.add(MaxPooling2D((2, 2)))

    model.add(Flatten())
    # model.add(Dense(128, acti vation='relu', kernel_initializer = 'he_normal', kernel_regularizer = l2(weight_decay)))
    # model.add(BatchNormalization())
    # model.add(Dropout(0.5))
    # output layer
    model.add(Dense(10, activation = 'softmax'))

    # optimize and compile model
    opt = Adam(learning_rate = 1e-3)
    model.compile(optimizer = opt, loss = 'categorical_crossentropy', metrics = ['accuracy'])

    return model

Answer 1

First get name of your desired layer using model.summary() .

Then use the name of that layer in place of desired_layer in below given code:

from keras.models import Model
extractor = Model(inputs=model.inputs, outputs=model.get_layer(desired_layer).output)
features = extractor.predict(x)

Here x is data from which you want to extract features.